JP7239798B2 - Element value inference method, element value inference device, and element value inference program - Google Patents

Description

The present invention relates to an element value inference method, an element value inference device, and an element value inference program.

In order to solve the problem "easily determining the constants of elements in a circuit" disclosed in Patent Document 1, for example, (1) a method of designing with a mathematical formula and a knowledge base, (2) a method of using a genetic program, and (3) a technique using a neural network is adopted.

JP-A-2004-145410

However, in method (1) above, whether or not the constants of the designed elements are appropriate depends on the designer's knowledge, experience, judgment, and the like. In addition, in the above method (2), since an attempt is made to obtain an appropriate solution through a genetic process (e.g., mutation, crossover of chromosomes), the same solution, that is, the same constant is repeated for the element. Lack of reproducibility when obtained. Furthermore, in the above method (3), the effect of learning can be expected only within the learned range. cannot adequately determine the constants of the elements.

SUMMARY OF THE INVENTION It is an object of the present invention to easily determine element values of circuit elements that improve circuit performance.

In order to solve the above-described problems, an element value inference method according to the present invention changes one element value from among a plurality of element values that can be taken by one or more circuit elements included in a circuit to another element value. an element value changing step; a simulation step of calculating a plurality of characteristics of the circuit using the other element values after the change; and a simulation step of calculating the characteristics of the circuit using the calculated characteristics. and a value calculation step of calculating the value of the element value of the one element value before the change, in order to infer an element value having a probability of improving the plurality of characteristics among the plurality of element values and an action-value function updating step of updating an action-value function indicating a correspondence relationship between the plurality of element values and the values of the plurality of element values, using the calculated values.

In order to solve the above-described problems, an element value inference apparatus according to the present invention changes one element value from among a plurality of element values that can be taken by one or more circuit elements included in a circuit to another element value. an element value changing unit; a simulation unit that calculates a plurality of characteristics of the circuit using the other element values after the change; and a value calculation unit for calculating the value of the element value of the one element value before the change, in order to infer an element value having a probability of improving the plurality of characteristics among the plurality of element values an action-value function updating unit that updates an action-value function indicating a correspondence relationship between the plurality of element values and the values of the plurality of element values, using the calculated values.

In order to solve the above problems, an element value inference program according to the present invention provides a computer with:
an element value changing step of changing from one element value to another element value among a plurality of element values that can be taken by one or more circuit elements included in a circuit, and using the other element value after the change , a simulation step of calculating a plurality of characteristics possessed by the circuit; a value calculation step of calculating a value of the one element value before the change using the calculated plurality of characteristics; Among them, in order to infer an element value having a probability of improving the plurality of characteristics, using the calculated value for the one element value before the change, the plurality of element values and the plurality of elements and an action-value function updating step of updating the action-value function indicating the correspondence between the value and the value.

According to the element value inference method, the element value inference device, and the element value inference program according to the present invention, the plurality of characteristics of the circuit calculated using the other element values after the change are used. calculates the value of the one element value before the change, and updates the action value function by using the value of the one element value before the change. By inferring element values that have a probability of improving the plurality of characteristics of the circuit using the updated behavioral value function, it is possible to easily determine element values of circuit elements that improve circuit performance. .

1 shows the configuration of an element value inference device of Embodiment 1. FIG. 2 shows the configuration (hardware) of the element value inference device of Embodiment 1. FIG. 1 shows the circuit of Embodiment 1; 1 shows a MOS field effect transistor of Embodiment 1. FIG. 4 shows the characteristics of the circuit of Embodiment 1. FIG. 2 shows a Q-table of Embodiment 1; The operation of the element value inference device of Embodiment 1 is shown (first half). The operation of the element value inference device of Embodiment 1 is shown (second half). 4 shows the updating of the Q-table of Embodiment 1; 4 shows circuit characteristics and values of the circuit characteristics of Embodiment 1. FIG. 1 shows the relationship between episodes, steps, and scores in Embodiment 1. FIG. 3 shows the relationship between learning time, step, and score in Embodiment 1. FIG. 3 shows a circuit configuration of a modified example; 4 shows circuit characteristics of a modified example and improvements in circuit characteristics.

<Embodiment 1>
An embodiment of an element value inference device according to the present invention will be described below.

<Configuration of element value inference device (outline)>
FIG. 1 shows the configuration of an element value inference device according to the first embodiment. The element value inference apparatus DV of the first embodiment uses reinforcement learning (Q-Learning), and uses MOS field effect transistors (MOSFETs), which are circuit elements that make up the circuit CK (shown in FIG. 3). The object is to infer "M" (described later with reference to FIG. 4), which is the element value of Oxide-Semiconductor Field-Effect Transistors M1 to M8. More specifically, the element value inference device DV has the probability of improving the circuit characteristics CH of the circuit CK, that is, the consumption current CH1 to the input conversion noise CH13 (shown in FIG. 5), in other words, the probability of approaching the overall optimum. “M” is inferred, that is, searched using an action-value function (hereinafter also referred to as “Q table QT”, shown in FIG. 6) for implementing Q-Learning.

Prior to the explanation of the element value inference device DV of the first embodiment, (1) the circuit CK, (2) the MOS field effect transistors M1 to M8 constituting the circuit CK, (3) the circuit characteristics CH of the circuit CK, and (4 ) Q table QT will be explained.

<Circuit configuration>
FIG. 3 shows the circuit of the first embodiment.

Circuit CK is a conventionally known operational amplifier circuit. The circuit CK, as shown in FIG. 3, includes MOS field effect transistors M1-M8, a resistor R1 and a capacitor C1 for differential amplification. A power supply voltage Vdd (eg, 5.0 V, 3.3 V) and a negative voltage Vss are applied to the circuit CK. The circuit CK performs the differential amplification on the two signals input to the terminal Vinm and the terminal Vinp, and outputs the differentially amplified signal from the terminal Vout.

<Structure of MOS field effect transistor>
FIG. 4 shows a MOS field effect transistor of Embodiment 1. FIG.

In the MOS type field effect transistors M1 to M8, as shown in FIG. 4, the source, drain and gate are formed on the substrate. "L" represents the gate length, which is the distance between the source and the drain, and "W" represents the gate width, which is the length of the gate in the depth direction. In the first embodiment, it is assumed that "L" and "W" are fixed and "W" is multiplied by "M" (M is an arbitrary integer). Here, the reason for trying to infer 'M' after fixing 'L' and 'W' is to shorten the time required for the entire learning by limiting the time required for the inference, that is, the search. It's for.

<Breakdown of circuit characteristics>
FIG. 5 shows circuit characteristics of the circuit of the first embodiment.

The circuit CK (shown in FIG. 3) has circuit characteristics CH, more specifically, current consumption CH1 to input conversion noise CH13, as shown in FIG. The values of the current consumption CH1 to the input equivalent noise CH13 change depending on the length of "W" of the MOS field effect transistors M1 to M8 that constitute the circuit CK, in other words, change depending on the magnitude of "M". Among the current consumption CH1 to the input equivalent noise CH13, if "M" is increased, there is a circuit characteristic in which the characteristic value changes to improve with respect to the minimum requirement RE. There are also circuit characteristics that change unfavorably with respect to . Therefore, it is necessary to infer "M" such that all of the current consumption CH1 to the input equivalent noise CH13 satisfy the minimum requirement RE and are optimal as a whole.

<Configuration of Q table>
FIG. 6 shows the Q table of the first embodiment.

The Q-table QT is a matrix, as shown in FIG. The vertical axis of the Q-table QT indicates the state S, and the horizontal axis indicates the action A. Values Q11 to Q66, for example, exist at positions (cells) defined by the vertical and horizontal axes.

"State S" is synonymous with "state" known for Q-Learning, and here refers to "circuit characteristics CH of circuit CK (consumption current CH1 to input conversion noise CH13)".

"Action A" is synonymous with "action" known for Q-Learning, and here, "in order to improve the circuit characteristics of the circuit CK (consumption current CH1 to input equivalent noise CH13), 'M' ( element value).

"Values Q11 to Q66" are synonymous with "Q values" known for Q-Learning. For example, "value Q23" is " Here, when the circuit characteristic CH (consumption current CH1 to input conversion noise CH13) of the circuit CK is in the "state S2", if the "M" (element value) of the circuit CK is changed, It means the value of "action A3" when taking "action A3" of

<Configuration of element value inference device (details)>
Returning to FIG. 1, as shown in FIG. 1, the element value inference device DV of Embodiment 1 includes a control unit 1, an initialization/initialization unit 2, an element value change unit 3, a simulation unit 4, A Q value calculator 5 , a Q table updater 6 , and a storage 7 are included.

The control unit 1 monitors and controls the overall operation of the element value inference device DV.

The initialization/initialization unit 2 initializes Q-Learning, initializes the Q table QT, sets the upper limit value (last number) of the step ST, sets the upper limit value (last number) of the episode EP, sets the target It is used for setting the value MK, setting α and γ of the update formula UD, and the like.

“Step ST” corresponds to the number of times “action A” (shown in FIG. 6) is changed, for example, “action A1” is performed at step ST=1, and “action A2” is performed at step ST=2. It means.

"Episode EP" refers to one set consisting of a plurality of steps ST. For example, when the episode EP is 3, for example, in the first episode, a plurality of steps ST of "action A1" → "action A2" → "action A3" → → "action A6" are performed, In the second episode, a plurality of steps ST of "action A1" → "action A3" → "action A5" →,,, → "action A6" are performed, and in the third episode, "action A1" → " A plurality of steps ST such as “action A4”→“action A2”→,,,→“action A6” are performed.

The target value MK is a constant used to calculate a reward r (a value for calculating a “value Q” (shown in FIG. 6)) represented by Equation 2, which will be described later. The value of the target value MK can be arbitrarily set by a person who uses the element value inference device DV. It is possible to set a score SC (Formula 1 described later) (before starting Q-Learning).

α and γ denote the learning rate and discount rate in the update formula UD represented by Formula 3, which will be described later, respectively. The learning rate α rate and the discount rate γ are known parameters for Q-Learning. Specifically, the learning rate α represents the degree to which the value Q is updated, and the discount rate γ represents the degree to which the future value Q is discounted. As with the target value MK described above, the person using the element value inference device DV can arbitrarily set the learning rate α and the discount rate γ.

The element value changer 3 changes the magnitude of "M", which is the element value of the MOS field effect transistors M1 to M8.

Every time the element value "M" is changed, the simulation unit 4 uses the changed "M" to calculate the values of the current consumption CH1 to the input equivalent noise CH13 of the circuit CK, that is, to simulate. . The simulation unit 4 performs the simulation using, for example, the conventionally known SPICE (Simulation Program with Integrated Circuit Emphasis). SPICE is based on a netlist (a plurality of theoretical equations EQ for calculating current consumption CH1 to input equivalent noise CH13) generated textually (characters) or graphically (image processing) based on the circuit CK. ” is used to calculate the values of the current consumption CH1 to the input conversion noise CH13.

The Q-value calculator 5 calculates "value Q", for example, "value Q11" to "value Q66" (shown in FIG. 6). Specifically, the Q value calculation unit 5 (1) calculates the score SC using the formula 1, (2) calculates the reward r using the formula 2, and (3) calculates the "value Q” is calculated. Here, Equation 1 is, for example, an evaluation criterion adopted in operational amplifier design contests. Equation 2 indicates that the score SC (evaluation value (synonymous with FIG. 10) to be described later) of "action A" when "action A" is performed is the target value MK (for example, the circuit characteristic CH of circuit CK). This is intended to give a reward r when the initial score (before learning) SC) is exceeded. The update formula UD in Equation 3 is a technique for updating the Q value known from Q-Learning.

・・・（式３） Score SC=(Slew rate CH6×In-phase input range CH11×DC gain CH3)/Power consumption CH2 (Formula 1)
Reward r=Score SC/Target value MK (Formula 2)

... (Formula 3)

Regarding Equation 3, originally in Q-Learning, t represents time, s _t means the state at time t (synonymous with state S (shown in FIG. 6)), and a _t is time It means an action at time t (synonymous with action A (illustrated in FIG. 6)), and A in Equation 3 means a collection of actions a in Equation 3. In the first embodiment, t represents the order in which "action A" (shown in FIG. 6) is performed. For example, when t=t1 (in other words, step ST=1), "action A3" is performed, When t=t2 (in other words, step ST=2), it represents that "action A5" is performed.

The Q table update unit 6 updates the Q table QT (illustrated in FIG. 6). Specifically, the Q table updating unit 6 updates the Q table QT by replacing the old “value Q” with the new “value Q” calculated by the Q value calculating unit 5 . For example, in the Q table QT, assuming that the "value Q23" of "behavior A3" in "state S2" was already "2", the Q value calculation unit 5 calculates "value Q23" as When "3" is newly calculated, the Q table updating unit 6 updates the Q table QT by rewriting the "value Q23" from the old "2" to the new "3".

The storage unit 7 stores information necessary for the control unit 1 to the Q table updating unit 6 to perform processing. As shown in FIG. 1, the storage unit 7 stores the circuit CK, the theoretical equation EQ, the circuit characteristic CH, the Q table QT, the updating equation UD, and the like.

<Configuration of Element Value Inference Device (Hardware)>
FIG. 2 shows the configuration (hardware) of the element value inference device of the first embodiment.

From the hardware point of view, the element value inference device DV includes an input unit 11, a CPU (Central Processing Unit) 12, an output unit 13, a storage medium 14, a memory 15, and including. The input unit 11 is composed of, for example, a keyboard and a mouse. The CPU 12 is the core of a well-known computer such as operating hardware according to software. The output unit 13 is composed of, for example, a liquid crystal monitor and a printer. The storage medium 14 stores the circuit CK, the theoretical equation EQ, the circuit characteristic CH, the Q table QT, the updating equation UD, etc., and further stores the program PR, like the storage unit 7 shown in FIG. . The storage medium 14 is composed of, for example, a hard disk drive (HDD: Hard Disk Drive), a solid state drive (SSD: Solid State Drive), and a ROM (Read Only Memory) in order to perform the storage. The memory 15 is composed of, for example, a DRAM (Dynamic Random Access Memory) and an SRAM (Static Random Access Memory).

Regarding the relationship between the function (FIG. 1) and the hardware (FIG. 2) in the element value inference device DV, the CPU 12 executes the program PR stored in the storage medium 14 using the memory 15 on the hardware. In addition, by controlling the operations of the input section 11 and the output section 13 as necessary, the functions of the control section 1 to the storage section 7 are realized.

<Operation of Element Value Inference Device>
7 and 8 are flowcharts showing the operation of the element value inference device of the first embodiment. The operation of the element value inference apparatus of Embodiment 1 will be described below with reference to FIGS. 7 and 8. FIG. "Step S" indicating the order of steps in FIGS. 7 and 8 has nothing to do with "Step ST", which is the number of times "action A" is changed.

Step S10: The user of the element value inference device DV designs the circuit CK, and more precisely, inputs information on the circuit CK to the element value inference device DV. Here, instead of newly designing the circuit CK, the user can adopt a conventionally known circuit (for example, an operational amplifier circuit which is one of the typical circuits in analog circuits). . The user inputs information about the circuit CK (for example, types of circuit elements and connection relationships between circuit elements) to the element value inference device DV in the form of characters or images using the mouse and keyboard as the input unit 11. to enter.

Step S11: The user of the element value inference device DV uses the initialization/initialization unit 2 to initialize Q-Learning. Specifically, the user determines how many times episode EP is performed (last number of episode EP), how many times step ST is performed (last number of step ST), target value MK, and learning rate α and the value of the discount rate γ. In order to facilitate the following explanation and understanding, it is assumed that the user sets "1000" as "last episode EP" and "90" as "last step ST". .

Step S12: The initialization/initialization unit 2 initializes the Q table QT. Specifically, the initial setting/initializing unit 2 sets an arbitrary value as the "value Q" of the Q table QT, in other words, randomly sets the "value Q". Here, the reason why the "value Q" of the Q table QT is set not to be blank but to some value is that if any "value Q" is blank, "behavior A1" to "A6" This is because it is not possible to determine which one of them should be selected preferentially. For example, it is desirable to use the widely known ε-greedy method so that the "value Q" does not depend on randomness.

Step S13: The initialization/initialization section 2 sets the episode EP to "0".

Step S14: The initialization/initialization unit 2 sets step ST to "0".

Step S15: The initial setting/initializing section 2 initializes the element value "M". More specifically, the initial setting/initializing unit 2 sets an arbitrary value (however, a fixed value that is common (same) among a plurality of episode EPs) as "M", for example, "2". Here, the reason for initializing "M" is that, for example, when the first episode EP ends, "M" is reset (initialized) to a fixed value, and then the second episode, which is the subsequent episode, is initialized. This is for starting the episode EP.

Step S16: The element value changer 3 changes the element value "M". The element value changing unit 3 changes "M" from "2" to "3", for example.

Step S17: The simulation unit 4 executes the theoretical formula EQ by substituting "M" which is "3" into the theoretical formula EQ described above.

Step S18: As a result of executing the theoretical formula EQ, the simulation unit 4 acquires the values of the circuit characteristics CH, that is, the values of the consumption current CH1 to the input conversion noise CH13.

Step S19: The simulation unit 4 calculates the values of the slew rate CH6, the in-phase input range CH11, the DC gain CH3), and the power consumption CH2 from among the obtained values of the current consumption CH1 to the input conversion noise CH13, and the values of the power consumption CH2 according to the above equation 1 , the "score SC" is calculated.

Step S20: The simulation unit 4 calculates the "reward r" by substituting the calculated "score SC" into the formula 2 described above.

Step S21: The simulation unit 4 substitutes the "reward r" calculated in step S20, the learning rate α and the discount rate γ set in step S11 into the above-described update formula UD (formula 3) to obtain " Calculate the value Q. Here, the calculated value Q is, for example, the value of the action A3 after considering the reward r that can be obtained by performing the action A3 in the state S2 and reaching the state S3. means More specifically, when the current consumption CH1 to the input conversion noise CH13 are all in the state S2, the calculated value Q is, for example, an action A3 of changing "M" to "3", and as a result, , the value of the action A3 in consideration of the reward r that can be obtained when all of the consumption current CH1 to the input-converted noise CH13 reach the state S3.

Step S22: The Q table updating unit 6 updates the Q table QT using the calculated "value Q" of the behavior A.

Step S23: The control unit 1 determines whether or not step ST is the last step "90".

Step S24: In step S23, when step ST is not "90", that is, when it does not reach "90", control unit 1 increments step ST by +1 and returns to step S16.

Step S25: In step S23, when step ST is "90", that is, when it reaches "90", the control unit 1 checks whether the episode EP is "1000" of the last episode. to decide.

Step S26: In step S25, when the episode EP is not "1000", that is, when it has not reached "1000", the control unit 1 increments the episode EP by +1 and returns to step S14.

In step S25, when the episode EP is "1000", that is, when it reaches "1000", the control unit 1 terminates the operation.

<Update Q table>
FIG. 9 shows the updating of the Q table of the first embodiment.

The Q table QT is updated as follows in steps S16 to S22 in the flow charts (FIGS. 7 and 8).

The element value changing unit 3 is 1. in FIG. , "action A3" means that "M" is set to "3" when the circuit characteristic CH (consumption current CH1 to input conversion noise CH13) of the circuit CK is in "state S2". can be performed, "action A3" is performed (corresponding to step S16 in FIG. 7). The simulation unit 4 determines that the circuit characteristic CH of the circuit CK corresponds to 2. in FIG. , it obtains that it will be in "state S3", and calculates the reward r using Equation 2 (corresponding to steps S17 in FIG. 7 to step S20 in FIG. 8). The Q value calculation unit 5 considers the value "5" of the "state S3" obtained by performing the "action A3" for the "value Q23" of the "action A3". , calculates "3" (corresponding to step S21 in FIG. 8), and the Q table updating unit 6 uses "3" in "Q23" to update the Q table QT (FIG. 8 corresponds to step S22 of ).

In the same way, the element value changing unit 3 performs 4. in FIG. , when the circuit characteristic CH of the circuit CK is in the "state S3" and the "action A4" meaning changing "M" to "4" can be performed, the "action A4 ” (corresponding to step S16 in FIG. 7). In the simulation unit 4, the circuit characteristic CH of the circuit CK corresponds to 5. in FIG. , it acquires that it will be in "state S4", and calculates the reward r using Equation 2 (corresponding to steps S17 in FIG. 7 to step S20 in FIG. 8). The Q value calculation unit 5 considers the value "6" of the "state S4" obtained by performing the "action A4" for the "value Q34" of the "action A4", and calculates 6. in FIG. , calculates "2" (corresponding to step S21 in FIG. 8), and the Q table updating unit 6 uses "2" in "Q34" to update the Q table QT (FIG. 8 corresponds to step S22 of ).

By repeating the updating of the Q table QT described above (corresponding to step S23 in FIG. 8), 7. in FIG. , the updated Q-table QT is completed. That is, until the step ST reaches "90", in other words, the value of "M" is changed "90" times. The Q table QT shown in is completed.

Regarding the relationship between the episode EP and the Q table QT, for example, the Q table QT completed by the end of the first episode EP is used when starting the subsequent second episode EP (FIG. 7, Arrow C) in FIG. In other words, the second episode EP starts with the Q-table completed in the preceding first episode EP.

Similarly, the third episode EP starts using the Q-table QT completed in the second episode EP, and the fourth episode EP starts using the Q-table QT completed in the third episode EP. be done.

In Embodiment 1, since the last number of episode EPs is set to "1000", one Q table QT is finally completed when the first episode EP to the 1000th episode EP are completed. to be completed.

Here, 7. in FIG. is completed when all of the 1st episode EP to the 1000th episode EP are completed. 7. in FIG. The Q-table QT shown in can also be completed when each episode from the 1st episode EP to the 1000th episode EP ends. Therefore, in one completed Q-table QT or in a plurality of completed Q-tables QT, it is inferred that the process of changing "M" to reach the largest "evaluation value" is optimal. do.

<Improvement of circuit characteristics>
FIG. 10 shows the circuit characteristics and values of the circuit characteristics of the first embodiment.

The evaluation value (synonymous with the score SC shown in Equation 1 above) for the circuit characteristic CH (consumption current CH1 to input conversion noise CH13) of the circuit CK (shown in FIG. 3) is as shown in FIG. , Before starting the above learning (FIGS. 7 and 8), it was 1.81×10 ¹⁹ , but after completing the above learning, it was changed to 7.71×10 ¹⁹ , That is, it is improved by about four times.

<Relationship between episodes, steps, and scores>
FIG. 11 shows the relationship between episodes, steps, and scores in Embodiment 1;

Of the "1000" episode EPs described above, as shown in FIG. SC is almost unchanged. From this, substantially, at the stage of about the "300th" episode EP, the score SC of steps ST from "1" to "90", that is, the circuit characteristic CH is improved to the maximum. In other words, if the episode EP is performed about "300" times, it is possible to obtain the reproducibility of repeatedly optimizing the circuit characteristics CH of the steps ST of "1" to "90". ing.

<Relationship between study period, step, and score>
FIG. 12 shows the relationship between learning time, step, and score in the first embodiment.

Regarding the above-described learning (FIGS. 7 and 8), in the initial stage of learning, as shown in FIG. The score SC of ST is lower than the pre-learning evaluation value (score SC) of "1.81×10 ¹⁹ ".

In the middle period of learning, as shown in FIG. × 10 ¹⁹ ”.

In the latter stage of learning, as shown in FIG. .81×10 ¹⁹ ”.

From the above three facts, as the learning progresses, the circuit characteristic CH of the circuit CK is improved from the above "1.81×10 ¹⁹ " (evaluation value (score SC) before learning). Gradual increase in ST is supported.

<Effect of element value inference device>
In the element value inference device DV of the first embodiment, as described above, the MOS field effect transistors M1 to M8 constituting the circuit CK are optimized in order to optimize the current consumption CH1 to the input conversion noise CH13 of the circuit CK as a whole. The value of M' is searched or inferred using Q-Learning. As a result, unlike the prior art, (1) there is no need to rely on the knowledge, experience, judgment, etc. of the designer who designs the circuit CK (FIGS. 7 and 8), and (2) the current consumption CH1 to the input conversion noise CH13 (3) Q-Learning itself limits the scope of learning Therefore, the value of "M" can be appropriately determined regardless of whether it is within or outside the learned range.

<Modification>
<Circuit configuration>
FIG. 13 shows the configuration of the circuit of the modification.

The modified circuit CK has a larger number of MOS field effect transistors M1 to M24 than the circuit CK of the first embodiment (shown in FIG. 3).

<Improvement of circuit characteristics>
FIG. 14 shows circuit characteristics and improvements in circuit characteristics of the modified example.

With respect to the modified circuit CK (shown in FIG. 13), the element value inference device DV (shown in FIGS. 1 and 2) performs the above-described learning (shown in FIGS. 7 and 8). , as shown in FIG. 14, the evaluation value (score SC) for the current consumption CH1 to the input-equivalent noise CH13 of the circuit CK of the modified example is 7.17×10 ¹⁸ before starting the learning described above. However, after completing the learning described above, it was changed to 1.26×10 ¹⁹ , that is, it was improved by about two times.

Since the circuit characteristics CH are improved in the first embodiment and the modification, it is expected that the circuit characteristics CH of other circuits (not shown) will also be improved through the above-described learning. be.

DV element value inference measure, 1 control unit, 2 initial setting/initialization unit, 3 element value change unit, 4 simulation unit, 5 Q value calculation unit, 6 Q table update unit, 7 storage unit

Claims (4)

the computer
an element value changing step of changing from one element value to another element value among a plurality of element values that can be taken by one or more circuit elements included in the circuit;
a simulation step of calculating a plurality of characteristics of the circuit using the other element values after the change;
A value calculation step of calculating the value of the one element value before the change using the calculated plurality of characteristics;
Among the plurality of element values, in order to infer an element value having a probability of improving the plurality of characteristics, using the calculated value for the one element value before the change, the plurality of element values and an action-value function updating step of updating an action-value function indicating the correspondence relationship between and the values of the plurality of element values;
An element value inference method that performs operations including The value calculation step rewards the calculation of the value of the one element value before the change when the calculated evaluation values regarding the plurality of characteristics exceed initial evaluation values regarding the plurality of characteristics of the circuit. 2. The element value inference method according to claim 1, wherein the element value inference method is performed by adding an element value changing unit that changes from one element value to another element value among a plurality of element values that can be taken by one or more circuit elements included in the circuit;
a simulation unit that calculates a plurality of characteristics of the circuit using the other element values after the change;
a value calculation unit that calculates the value of the one element value before the change using the plurality of calculated characteristics;
Among the plurality of element values, in order to infer an element value having a probability of improving the plurality of characteristics, using the calculated value for the one element value before the change, the plurality of element values and an action-value function updating unit that updates an action-value function indicating a correspondence relationship between and the values of the plurality of element values;
Element value reasoning device including. to the computer,
an element value changing step of changing from one element value to another element value among a plurality of element values that can be taken by one or more circuit elements included in the circuit;
a simulation step of calculating a plurality of characteristics of the circuit using the other element values after the change;
A value calculation step of calculating the value of the one element value before the change using the calculated plurality of characteristics;
Among the plurality of element values, in order to infer an element value having a probability of improving the plurality of characteristics, using the calculated value for the one element value before the change, the plurality of element values and an action-value function updating step of updating an action-value function indicating the correspondence relationship between and the values of the plurality of element values;
Element value inference program to run.

Images